@wastech-mdlint/mcp-server
v0.2.1
Published
MCP server for Markdown linting: read-only Model Context Protocol tools for lint findings, the context graph, and change impact.
Readme
@wastech-mdlint/mcp-server
A Model Context Protocol server that gives an agent read-only access to a repository's Markdown context: lint findings, the context graph, reachability slices, change impact, and a compiled skill.
stdio only, read-only, six tools. No HTTP or SSE transport, no mutating tools, no code-plugin execution. Every tool runs through @wastech-mdlint/core, the same pipeline the CLI uses — this package is the host adapter around it: tool registration, input validation, structured output, and error wrapping.
Run it
npx @wastech-mdlint/mcp-server # run directly, no install
npm i -D @wastech-mdlint/mcp-server # or install the wastech-mdlint-mcp binAdd it to any stdio-based MCP host (Claude Code's .mcp.json, Claude Desktop's claude_desktop_config.json, and so on):
{
"mcpServers": {
"wastech-mdlint": {
"command": "npx",
"args": ["-y", "@wastech-mdlint/mcp-server"],
},
},
}Node.js >=24.17.0. Readiness is announced on stderr; stdout carries only the protocol, so nothing the server logs can corrupt a JSON-RPC frame.
The six tools
| Tool | What it does |
| --- | --- |
| lint | Lint ad-hoc Markdown content against an explicit rule list. Does not load project config. |
| lint-files | Lint the project's Markdown files using the resolved config, or the zero-config default. |
| context-graph | Build the project's context graph — nodes, edges, cycles, or a summary document. |
| context-slice | The files reachable within depth hops of a query resolved by exact match. |
| impact-analysis | The blast radius of changing one file: direct and transitive dependents. |
| compile-context | Compile the project skill (SKILL.md) from config.compile. |
All six carry a readOnlyHint annotation. Five return structuredContent alongside an outputSchema; compile-context returns two plain-text content blocks instead.
The five file-based tools accept a cwd. A cwd that does not exist, or that names a file rather than a directory, is rejected with an actionable error rather than answered with an empty result — an empty corpus is indistinguishable from a clean repository, so a plausible-looking success would answer a different question than the one asked. Omit it to analyze the server process's own working directory.
Errors
Tool failures use a structured { code, message, hint } contract, carried in structuredContent with message and hint also rendered into the text block, so a host that shows only text still sees the actionable half.
One limit is deliberate: that contract covers failures the tool itself detects. An argument shape the tool's advertised inputSchema rejects outright is refused by the protocol layer before the tool runs, and comes back as a validation error with no structuredContent at all.
Documentation
- MCP server guide — every tool's arguments, the full error contract, and host wiring
- Project README · Configuration · Output & exit codes
License
MIT — see LICENSE.
